|This article does not cite any references or sources. (November 2010)|
A memory barrier, also known as a membar, memory fence or fence instruction, is a type of barrier instruction which causes a central processing unit (CPU) or compiler to enforce an ordering constraint on memory operations issued before and after the barrier instruction. This typically means that certain operations are guaranteed to be performed before the barrier, and others after.
Memory barriers are necessary because most modern CPUs employ performance optimizations that can result in out-of-order execution. This reordering of memory operations (loads and stores) normally goes unnoticed within a single thread of execution, but can cause unpredictable behaviour in concurrent programs and device drivers unless carefully controlled. The exact nature of an ordering constraint is hardware dependent and defined by the architecture's memory ordering model. Some architectures provide multiple barriers for enforcing different ordering constraints.
Memory barriers are typically used when implementing low-level machine code that operates on memory shared by multiple devices. Such code includes synchronization primitives and lock-free data structures on multiprocessor systems, and device drivers that communicate with computer hardware.
An illustrative example 
When a program runs on a single-CPU machine, the hardware performs the necessary bookkeeping to ensure that the program executes as if all memory operations were performed in the order specified by the programmer (program order), so memory barriers are not necessary. However, when the memory is shared with multiple devices, such as other CPUs in a multiprocessor system, or memory mapped peripherals, out-of-order access may affect program behavior. For example, a second CPU may see memory changes made by the first CPU in a sequence which differs from program order.
The following two-processor program gives an example of how such out-of-order execution can affect program behavior:
Initially, memory locations
f both hold the value
0. The program running on processor #1 loops while the value of
f is zero, then it prints the value of
x. The program running on processor #2 stores the value
x and then stores the value
f. Pseudo-code for the two program fragments is shown below. The steps of the program correspond to individual processor instructions.
while (f == 0); // Memory fence required here print x;
x = 42; // Memory fence required here f = 1;
One might expect the print statement to always print the number "42"; however, if processor #2's store operations are executed out-of-order, it is possible for
f to be updated before
x, and the print statement might therefore print "0". Similarly, processor #1's load operations may be executed out-of-order and it is possible for
x to be read before
f is checked, and again the print statement might therefore print an unexpected value. For most programs neither of these situations are acceptable. A memory barrier can be inserted before processor #2's assignment to
f to ensure that the new value of
x is visible to other processors at or prior to the change in the value of
f. Another can be inserted before processor #1's access to
x to ensure the value of
x is not read prior to seeing the change in the value of
For another illustrative example (a non-trivial one that arises in actual practice), see double-checked locking.
Low-level architecture-specific primitives 
Memory barriers are low-level primitives and part of an architecture's memory model, which, like instruction sets, vary considerably between architectures, so it is not appropriate to generalize about memory barrier behavior. The conventional wisdom is that using memory barriers correctly requires careful study of the architecture manuals for the hardware being programmed. That said, the following paragraph offers a glimpse of some memory barriers which exist in contemporary products.
Some architectures, including the ubiquitous x86/x64, provide several memory barrier instructions including an instruction sometimes called "full fence". A full fence ensures that all load and store operations prior to the fence will have been committed prior to any loads and stores issued following the fence. Other architectures, such as the Itanium, provide separate "acquire" and "release" memory barriers which address the visibility of read-after-write operations from the point of view of a reader (sink) or writer (source) respectively. Some architectures provide separate memory barriers to control ordering between different combinations of system memory and I/O memory. When more than one memory barrier instruction is available it is important to consider that the cost of different instructions may vary considerably.
Multithreaded programming and memory visibility 
Multithreaded programs usually use synchronization primitives provided by a high-level programming environment, such as Java and .NET Framework, or an application programming interface (API) such as POSIX Threads or Windows API. Primitives such as mutexes and semaphores are provided to synchronize access to resources from parallel threads of execution. These primitives are usually implemented with the memory barriers required to provide the expected memory visibility semantics. In such environments explicit use of memory barriers is not generally necessary.
Each API or programming environment in principle has its own high-level memory model that defines its memory visibility semantics. Although programmers do not usually need to use memory barriers in such high level environments, it is important to understand their memory visibility semantics, to the extent possible. Such understanding is not necessarily easy to achieve because memory visibility semantics are not always consistently specified or documented.
Just as programming language semantics are defined at a different level of abstraction than machine language opcodes, a programming environment's memory model is defined at a different level of abstraction than that of a hardware memory model. It is important to understand this distinction and realize that there is not always a simple mapping between low-level hardware memory barrier semantics and the high-level memory visibility semantics of a particular programming environment. As a result, a particular platform's implementation of (say) POSIX Threads may employ stronger barriers than required by the specification. Programs which take advantage of memory visibility as-implemented rather than as-specified may not be portable.
Out-of-order execution versus compiler reordering optimizations 
Memory barrier instructions only address reordering effects at the hardware level. Compilers may also reorder instructions as part of the program optimization process. Although the effects on parallel program behavior can be similar in both cases, in general it is necessary to take separate measures to inhibit compiler reordering optimizations for data that may be shared by multiple threads of execution. Note that such measures are usually only necessary for data which is not protected by synchronization primitives such as those discussed in the prior section.
In C and C++, the volatile keyword was intended to allow C and C++ programs to directly access memory-mapped I/O. Memory-mapped I/O generally requires that the reads and writes specified in source code happen in the exact order specified with no omissions. Omissions or reorderings of reads and writes by the compiler would break the communication between the program and the device accessed by memory-mapped I/O. A C or C++ compiler may not reorder reads and writes to volatile memory locations, nor may it omit a read or write to a volatile memory location. The keyword volatile does not guarantee a memory barrier to enforce cache-consistency. Therefore the use of "volatile" alone is not sufficient to use a variable for inter-thread communication on all systems and processors
The C and C++ standards prior to C11 and C++11 do not address multiple threads (or multiple processors), and as such, the usefulness of volatile depends on the compiler and hardware. Although volatile guarantees that the volatile reads and volatile writes will happen in the exact order specified in the source code, the compiler may generate code (or the CPU may re-order execution) such that a volatile read or write is reordered with regard to non-volatile reads or writes, thus limiting its usefulness as an inter-thread flag or mutex. Preventing such is compiler specific, but some compilers, like gcc, will not reorder operations around in-line assembly code with volatile and "memory" tags, like in: asm volatile ("" : : : "memory"); (See more examples in compiler memory barrier). Moreover, it is not guaranteed that volatile reads and writes will be seen in the same order by other processors or cores due to caching, cache coherence protocol and relaxed memory ordering, meaning volatile variables alone may not even work as inter-thread flags or mutexes.
Some languages and compilers may provide sufficient facilities to implement functions which address both the compiler reordering and machine reordering issues. In Java version 1.5 (also known as version 5), the volatile keyword is now guaranteed to prevent certain hardware and compiler re-orderings, as part of the new Java Memory Model. C++11 standardizes special atomic types and operations with semantics similar to those of volatile in the Java Memory Model.
See also 
- Memory Barriers: a Hardware View for Software Hackers
- Microsoft Driver Development: Memory Barriers on Multiprocessor Architectures
- HP technical report HPL-2004-209: Threads Cannot be Implemented as a Library
- Linux kernel memory barrier issues on multiple types of CPUs
- Documentation on memory barriers in the Linux kernel
- Handling Memory Ordering in Multithreaded Applications with Oracle Solaris Studio 12 Update 2: Part 1, Compiler Barriers
- Handling Memory Ordering in Multithreaded Applications with Oracle Solaris Studio 12 Update 2: Part 2, Memory Barriers and Memory Fences